[bugfix] Resolve memory leak for evaluation (#6326)

* resolve bug

* resolve flake8

* revert name

Author: tchaton

CHANGELOG.md

@@ -39,6 +39,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Check `LightningOptimizer` doesn't delete optimizer hooks ([#6305](https://github.com/PyTorchLightning/pytorch-lightning/pull/6305)
 
 
+- Resolve memory leak for evaluation ([#6326](https://github.com/PyTorchLightning/pytorch-lightning/pull/6326)
+
+
 ## [1.2.2] - 2021-03-02
 
 ### Added

pytorch_lightning/trainer/evaluation_loop.py

@@ -203,6 +203,10 @@ def __run_eval_epoch_end(self, num_dataloaders):
 
         # with a single dataloader don't pass an array
         outputs = self.outputs
+
+        # free memory
+        self.outputs = []
+
         eval_results = outputs
         if num_dataloaders == 1:
             eval_results = outputs[0]

tests/trainer/logging_/test_eval_loop_logging_1_0.py

@@ -126,6 +126,7 @@ def validation_step_end(self, acc):
         def validation_epoch_end(self, outputs):
             self.log('g', torch.tensor(2, device=self.device), on_epoch=True)
             self.validation_epoch_end_called = True
+            assert len(self.trainer.evaluation_loop.outputs) == 0
 
         def backward(self, loss, optimizer, optimizer_idx):
             return LightningModule.backward(self, loss, optimizer, optimizer_idx)